A Global Registration Method Based on the Vector Field Representation

In this paper, we propose a new approach for point cloud registration based on a volumetric representation called the vector field. A surface classification method is first integrated in the vector field represention to implement rough registration estimation in a pair-wise manner. A pose refinement process is then applied to the rough-estimate. The global registration process is supported entirely by the vector field representation. It does not require a priori information on relative position of views and reduces the computational complexity for searching correspondences. Experiments demonstrate the efficiency of the method on a wide variety of objects.